Polarization Methods in Machine and Computer Vision

该项目扩展了基于偏振的视觉传感在低级机器视觉和高级计算机视觉中的独特能力。这篇由首席研究员撰写的博士论文开辟了视觉研究的一个新的子领域，利用反射偏振分析作为一种新的感官介质来提取目标特征，它在解决一些重要的图像理解问题上具有现有的基于强度和颜色的技术的关键优势。利用这些方法的基本基础，该研究探索了各种基于偏振的技术的发展，用于机器视觉中的检测过程，以及计算机视觉中的对象识别。反射极化分析提供了视觉上第一个实用的方法，用于根据不同的电导率水平区分和分类材料表面，金属和介质的表面处于极端状态。对于更受控制的机器视觉环境，人们研究了基于偏振的方法，这些方法显著改进了现有的方法，用于准确地无损检测材料的组成和缺陷，如裂纹和划痕。一个潜在的非常有用的应用是对VLSI芯片的检查。这项研究还将研究如何将基于偏振的材料分类方法用于计算机视觉，以根据用于物体识别的子部件的材料组成来增强物体描述。基于偏振的方法可以非常简单地用于提取形状约束；该项目研究了如何将其用于物体识别。最后，研究探索了基于偏振的方法在颜色恒定中的应用，以同时确定物体和光源的颜色。

This project extends the unique capabilities of polarization- based visual sensing both in low-level machine vision and "higher-level" computer vision. The doctoral thesis written by the principle investigator has opened up a new subarea of vision research utilizing reflected polarization analysis as a new sensory medium for object feature extraction which has been demonstrated to have key advantages over existing intensity and color-based techniques for some important image understanding problems. Utilizing the fundamental foundations for these methods the research explores the development of a variety of polarization-based techniques for inspection processes in machine vision, and object recognition in computer vision. Reflected polarization analysis has provided the first practical methodology in vision for distinguishing and classifying material surfaces according to different levels of electrical conductivity, with metals and dielectrics at the extremes. For the more controlled environments of machine vision, polarization-based methods are studied that significantly improve upon existing methods for accurate non-destructive inspection of materials in terms of their composition and defects such as cracks and scratches. One potentially very useful application is to inspection of VLSI chips. The research will also study how material classification from polarization-based methods can be used in computer vision to augment object descriptions in terms of material composition of subparts for object recognition. Polarization-based methods can be very simply used to extract shape constraints; the project examines how this can be used for object recognition. Finally, the research explores the application of polarization-based methods to color constancy, for simultaneous determination of object and light source color.